The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill builds agents that ingest untrusted user text which is then processed by an LLM to generate executable code.
Ingestion points: Text inputs to methods such as analyze_feedback(text), triage(message), handle_inquiry(question), and answer_question(question) in SKILL.md.
Boundary markers: Absent; the examples demonstrate prompts derived from method docstrings without explicit delimiters or "ignore instructions" guards for the input data.
Capability inventory: The agents possess arbitrary Python code execution capabilities within a REPL environment, including access to self and library imports.
Sanitization: The skill documents the use of AST validation and module deny-lists but explicitly notes these are not complete containment boundaries and recommends running in a sandboxed environment.
[DYNAMIC_EXECUTION]: The NOOA framework's primary mechanism involves the LLM writing and executing Python code at runtime to implement "generation methods" (methods with ... bodies).
[COMMAND_EXECUTION]: The skill provides instructions for installing packages using pip and uv, as well as commands for starting a development trace viewer (uv run nooa start-dev) which listens on a local network port.
[EXTERNAL_DOWNLOADS]: The skill directs the user to download the framework from NVIDIA's official GitHub repository (NVIDIA-NeMo/labs-OO-Agents).

nvidia-oo-agents